Satellite image time series (SITS) are crucial for monitoring vegetation dynamics and land use change. In this study, we investigated various aspects of SITS processing and analysis, focusing on the task of grassland and crop classification. We explored the optimal SITS density for grassland classification and found that three to four observations per month are sufficient to achieve results that differ by less than 0.05 in the F1 score from those obtained using full cloud-free SITS. We examined the possibility of early crop classification across different years using transfer learning. Models trained only on source years achieved an average F1 score of 0.825 (threshold value) on target years. Using transfer learning, we surpassed the F1 threshold value with 48,000 training samples (6% of all available reference data) from the target year as early as mid-July. We introduced a novel Radar-Optical Vegetation Index (ROVI) that combines the Normalized Difference Vegetation Index (NDVI) and radar coherence, enabling the creation of denser time series and achieving better results than traditional gap-filling methods. Our research contributes significantly to a better understanding of SITS processing and analysis and provides guidance for effective vegetation monitoring, even under conditions with limited reference data. The evaluation of procedures and methods allows for informed method selection, as our study provides insight into the results of their application for grassland and crop classification. Furthermore, our study demonstrates the potential of combining radar and optical data to improve the reliability and density of time series for monitoring development of vegetation.
|
